AI papers — 2026-09-14

Problem: Open-weight post-training for frontier cyber-agent capability is bottlenecked less by model scale than by the cost of executable environments, reliable multi-turn supervision, and access to strong teacher models.

Method: A seven-person independent team post-trains three open-weight checkpoints (Qwen3.6-35B-A3B, Qwen3.8-27B, Qwen3.5-122B-A10B) via SFT on 164,269 execution-verified trajectories from an environment-grounded data engine spanning repository-level coding, CVE reproduction, CTF, Linux-kernel-history mining, exploit development, firmware, and physical-device environments. The pipeline is supported by five auxiliary techniques: Choulea (recovering encrypted hidden-reasoning traces from closed frontier APIs), SkyReal (reducing teacher-sampling cost via low-cost account arbitrage), Hongzwang (a jailbreak/retry system for eliciting restricted teacher completions), PSBreakup (white-box reverse-KL distillation to restore cyber capability weakened by model merging), and Kreator (rewriting human-expert interventions into teacher-native SFT trajectories for tasks no available teacher can solve alone).

Key results:

• Average CyberGym verified-success improves 23.76% and pooled CTF success improves 10.49% over base checkpoints.
• Feyospace-s1 reaches 63.24% verified success and ranks 10th on the official CyberGym leaderboard (as of Sept 1, 2026); all three checkpoints rank 1st among models of comparable parameter scale.
• The S7 identity-renaming jailbreak strategy resolves an additional 39% of otherwise-unsolved elicitation cases.
• PSBreakup recovers merge-weakened cyber capability across DeepSeek-V4-Pro, GLM-5.2, and Kimi K3.

Why it matters / caveats: The paper is notable — and concerning — for openly documenting techniques adjacent to abuse: large-scale gray-market account-harvesting to cheaply access competitor APIs (SkyReal), systematic jailbreaking of closed frontier models to elicit exploit-development content (Hongzwang), and extraction of competitors' encrypted hidden reasoning traces (Choulea). The authors state SkyReal, Hongzwang, and the recovered reasoning traces were not used in the final training mixture (excluded for cost/quality reasons), but the detailed disclosure of these methods, plus real V8/kernel exploit-development trajectories, raises dual-use and responsible-disclosure concerns.

Problem: Robot foundation models achieve strong in-distribution performance but generalize poorly under visual distribution shifts because action generation can exploit task-irrelevant, training-distribution-specific visual correlations ("vision–action shortcuts").

Method: Latent Interface Training (LIT), a model-agnostic two-stage recipe. Stage 1 trains the action expert from scratch on language, robot state, and each demonstrated chunk's terminal SE(3) end-effector pose (no images), learning a spatial-goal-conditioned action prior. Stage 2 introduces vision through a pose-supervised "latent interface": learnable latent tokens cross-attend to visual/semantic backbone features and serve as the action expert's only visual conditioning path, supervised to reconstruct the same terminal pose used in Stage 1.

Key results:

• Across four VLA/WAM architectures (π0.5, MolmoAct2, FAST-WAM, ImageWAM), LIT improves overall LIBERO-Plus success by 3.87–10.70 percentage points while preserving/improving average LIBERO success.
• Real-world evaluations show 13.30–16.70 percentage-point gains aggregated across three manipulation tasks under unseen camera configurations, lighting variation, and distractors.

Why it matters / caveats: Offers a general, architecture-agnostic recipe to reduce shortcut learning in robot foundation models without redesigning the backbone. Evaluated on simulation (LIBERO-Plus) plus a limited set of real-world tasks, so generalization to larger-scale real-world robot fleets remains to be shown.

Problem: Researchers lack a single place to discover relevant AI benchmarks, locate their datasets/code, and understand the evaluation settings behind reported scores.

Method: Benchmark Radar, a living database and search engine that combines daily discovery (37 sources: 13 direct connectors + 24 first-party feeds, with arXiv, Hugging Face Hub, and GitHub Search as core sources) with a searchable catalog built from four sources (LLM Stats, OpenCompass Hub, Artificial Analysis, model reports), using BM25F lexical search and reviewed identity links to merge related records while preserving separate measurements. Ships as a web dashboard with a leaderboard, Pareto-frontier view, saturation/trend views, and an offline CLI.

Key results:

• The v0.11.0 catalog contains 1,283 source records across 4 catalog sources.
• 12,916 numeric score observations recorded on 790 records (493 records currently lack numeric scores).
• Daily discovery draws on 37 total sources.

Why it matters / caveats: A useful meta-tool for benchmark discovery and provenance-tracking as the number of LLM evaluations proliferates. The paper is explicitly an infrastructure/census contribution rather than a new capability result, and its coverage and quality depend on the health of the external source feeds it aggregates.

Problem: Trainable post-training attention-sparsification methods rely on non-differentiable hard Top-K context selection, forcing them to use auxiliary dense-attention distillation that creates a "ranking misalignment" between what the selector learns and what actually helps final predictions under a fixed attention budget.

Method: Simple Attention Sparsification (SAS), a gated sparse attention mechanism that injects the selector's continuous scores as log-space gates directly into the qKᵀ attention logits (inside the softmax), letting the language-modeling loss backpropagate through the selector end-to-end with no auxiliary distillation. Key design choices: log-space gate placement, softmax-normalized gates, retaining continuous (non-binarized) scores, and updating only the selected blocks during training. A custom Triton kernel fuses gate addition into FlashAttention-style tiled computation for efficient long-context training.

Key results:

• At a 1024-token attention budget, SAS improves over SeerAttention-R by 6.0–7.7 points on MATH500 and 10.6–15.5 points on GPQA-Diamond (across Qwen3-4B/8B/14B).
• On LongBench, SAS leads at every budget (e.g., +3.2 on the 8K+ bucket at budget 2048 for Qwen3-4B).
• On agentic tasks, it improves BFCL by up to +3.5 and nearly recovers full-attention performance on VitaBench at budget 4096.

Why it matters / caveats: Removes the need for teacher-attention distillation in trainable sparse attention, simplifying training while improving accuracy especially under tight budgets. Gains are largest at low budgets, and the method is demonstrated as a post-training/mid-training adaptation of existing dense checkpoints rather than from-scratch pretraining at very large scale.

Problem: Existing LLM-agent skill-optimization methods rely on costly execution-based evaluation and substantial task data to refine reusable skills, making optimization expensive.

Method: COBRA-Skills formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space, combining contextual-bandit-guided prioritization (a neural reward predictor plus a LinearUCB-style uncertainty bonus over skill embeddings) with evidence-grounded evolutionary operators — regeneration (new skills from no-skill trajectories), rollout mutation (revising the evaluated skill using its own successful/failed trajectories), and crossover (combining a strong backbone skill with another strong skill as positive evidence and weak skills as negative evidence) — periodically refreshing the skill population.

Key results:

• Across six heterogeneous agent benchmarks and three target models (including Qwen3.6-35B-A3B, GPT-5.4-Nano, Gemma4-26B-A4B-it), COBRA-Skills achieves the strongest average performance among compared methods.
• Reduces optimization cost by 55–58% relative to SkillOpt, using only 50 unique optimization examples per benchmark.

Why it matters / caveats: Makes skill optimization substantially cheaper by focusing evaluation budget on promising/uncertain candidates rather than exhaustive execution-based search. Robustness is shown across agent harnesses and when the target model itself generates/refines skills, though results are reported as cross-benchmark averages rather than uniform per-benchmark wins.

Problem: Existing audio-generation approaches specialize separately for speech, sound effects, and music with incompatible representations, and unifying them in a discrete, LLM-compatible framework risks catastrophic interference with the LLM's inherited text capabilities.

Method: StepAudio 3 Gen is a discrete autoregressive generator built on a 12.5 Hz StepAudio Tokenizer that jointly quantizes semantic and acoustic features into a shared 16×2048 residual-vector-quantization (RVQ) code space. The LLM backbone autoregressively predicts the first codebook along time, while a lightweight 4-layer causal Transformer predicts the remaining 15 codebooks along the depth axis (a time–depth architecture). A zero-initialized, token-wise "RVQ Adaptor" reconciles multi-codebook embeddings with the pretrained text embedding space, and a four-stage interference-aware progressive pretraining curriculum (frozen-backbone alignment, joint understanding with 50% text replay, gradient-detached generation, end-to-end cool-down) preserves textual abilities.

Key results: Achieves state-of-the-art performance on both TTS and voice design among compared systems while retaining strong generation across speech, vocals, sound effects, and music. (Detailed numeric benchmark tables appear later in the paper and were not fully reviewed for this summary.)

Why it matters / caveats: A fully discrete, diffusion-free unified audio generator that keeps text and audio in one causal LLM stream is architecturally distinct from the diffusion/flow-based unified-audio trend, offering a path to jointly optimize text and audio intelligence. Full quantitative comparisons against other unified audio models (e.g., UniAudio 2.0) were Not stated in the portion reviewed.

Problem: Direct Preference Optimization assumes every observed preference label is reliable, but real preference data often contains reversed, weak, or ambiguous labels that cause harmful policy updates.

Method: Posterior Label Correction DPO (PLC-DPO) treats each observed pair's direction as a latent clean/flip/tie state and estimates a posterior-like routing distribution from the standardized (EMA-calibrated) policy-reference preference margin. It stops gradients through the routing weights and uses them to softly mix forward DPO, reversed DPO, and a tie-regularizing loss, with a warm-up schedule and confidence-gated correction strength to avoid early-training instability.

Key results:

• Across 57 dataset–model–benchmark cells, PLC-DPO obtains the best mean win rate against DPO (60.5 vs. 55.5 for the next-best method).
• Shows stable routing that distinguishes flipped from weakly-directional pairs under injected-noise/tie stress tests and an independent human-disagreement evaluation set.

Why it matters / caveats: Directly corrects noisy pair labels online rather than merely filtering or uniformly smoothing them, requiring no additional supervision or reward model. The correction signal relies on the policy-reference margin as a proxy for label reliability, so effectiveness may depend on how well-calibrated that margin is early in training.

Problem: Many proposed RLVR "data policy" interventions (rollout selection, reweighting, adaptive domain mixtures) are evaluated inconsistently, so it's unclear whether reported gains over uniform sampling are reproducible once training recipe, seeds, and evaluation coverage are controlled.

Method: DataFlex-RL is an evaluation platform that isolates selection, reweighting, and mixture-adaptation interventions from the rest of a fixed GRPO training recipe, using a standardized math/logic/science corpus, matched seeds, and a common 12-benchmark evaluation harness. The primary experiment compares 13 configurations across 12 matched seeds on Qwen2.5-7B-base (591 total runs in the release).

Key results:

• Uniform GRPO improves domain-balanced accuracy by 7.76 points over the untrained checkpoint.
• None of 8 selection/reweighting methods, nor 3 adaptive mixtures, shows a paired 95% CI excluding zero relative to uniform sampling/fixed mixture (mean spreads of only 0.97 and 0.61 points respectively).
• A 12-seed Llama-3.1-8B-base extension finds no common winner among methods.
• Evaluation-summary choice matters: rankings from a math-heavy 6-benchmark summary and the full domain-balanced 12-benchmark summary are negatively correlated (ρ = −0.33).

Why it matters / caveats: A sobering negative result: under controlled seeds and full benchmark coverage, none of the tested RLVR data-selection/reweighting/mixture tricks reliably beats uniform sampling, and apparent "wins" in prior papers may stem from narrow evaluation-benchmark coverage rather than real gains — a caution for the RLVR literature broadly.

Problem: Adaptively routing prompts to the best of several LLM "experts" to maximize response quality is hard because response quality (reward) is rarely observed, and obtaining feedback (human or LLM-judge) is expensive and thus limited to a budget m ≪ T rounds.

Method: The problem is formulated as a linear contextual bandit over K expert arms and d-dimensional prompt-embedding contexts, where the algorithm can adaptively choose which of at most m rounds to request feedback for. In the full-information setting (LimFullFeed), feedback is requested at regular intervals chosen to maximize the determinant of the covariance matrix (information gain); the bandit-setting extension maintains separate per-expert confidence sets to preserve regret guarantees, and a further extension (in the appendix) handles varying per-expert feedback costs.

Key results:

• Full-information regret of Õ(dT/√m); bandit-setting regret of Õ(dT√(K/m)).
• Experiments on the Nectar dataset (GPT-3.5, GPT-3.5-Turbo-Instruct, GPT-4, GPT-4-0613, LLaMA-2-7B-Chat, Mistral-7B-Instruct) show the algorithms learn high-quality routing policies from limited feedback, and periodic feedback selection performs nearly as well as hindsight-optimal choices.

Why it matters / caveats: Provides theoretically-grounded regret bounds for a practically important, underexplored setting — LLM routing under a hard cap on how often quality feedback can be collected — relevant for reducing human/judge-evaluation costs in production routing systems. Results are demonstrated on one preference dataset (Nectar) with a modest set of older LLMs, so scaling to larger/more diverse expert pools is untested here.

Problem: Existing part-aware 3D asset generation methods can produce visually complete individual parts without ensuring the parts form a valid, stable physical assembly — neighboring parts can interpenetrate, lack real connections, or collapse under gravity.

Method: A three-stage, physics-guided framework applied on top of an off-the-shelf image-to-3D part generator (Hunyuan3D reconstruction + X-Part decomposition): (1) collision-free part geometry editing that trims overlapping volume via a pairwise intrusion score, (2) connector geometry reasoning that builds a contact graph from proximity/facing-normal criteria and initializes parameterized peg-and-socket connectors at each contact surface, and (3) physics-based connector optimization that refines connector position, direction, and scale using simulation feedback (Incremental Potential Contact for intersection-free rigid-body contact) via an evolutionary search over connector parameters.

Key results:

• Achieves 95.0% assembly stability in simulation versus nearly 0% for existing part-aware 3D generation baselines.
• Remains competitive on geometric fidelity metrics (Chamfer Distance, F1-score).
• Results validated further through 3D printing and manual real-world assembly.

Why it matters / caveats: First method to directly target physical assembly validity (not just visual completeness) for single-image part-aware 3D generation, enabling fabrication-ready outputs. Evaluated on a subset of one benchmark (HY3D-Bench) and builds on a specific upstream part generator, so generality across other part-decomposition backbones is Not stated.

Problem: Routing user requests over large LLM-agent skill registries (tens of thousands of skills) with independent top-k relevance ranking wastes limited context budget on functionally redundant skills, especially for complex tasks needing complementary skill sets.

Method: Diverse Skill Routing (DSR), a diversity-aware reranking framework built on a standard retrieve-and-rerank pipeline: an encoder retriever narrows the candidate pool, a pointwise reranker assigns query-dependent quality scores, and a Determinantal Point Process (DPP) with a novel "query-residual diversity kernel" — which removes each skill's query-aligned component before computing inter-skill similarity — performs greedy MAP selection to balance relevance against true redundancy (rather than penalizing skills merely for sharing query relevance).

Key results:

• On the SkillRouter benchmark (~80K candidate skills), DSR improves recall and full coverage over a strong pointwise SkillRouter baseline, with larger gains on multi-skill queries and at larger retrieval cutoffs.
• Ablations show replacing the query-residual kernel with a standard inter-skill similarity kernel substantially reduces multi-skill full coverage.

Why it matters / caveats: Reframes large-scale skill routing as complementary-set selection rather than pure relevance ranking, which should matter increasingly as agent skill libraries scale into the tens of thousands. Evaluated on a single benchmark (SkillRouter), so transfer to other skill/tool registries and end-to-end agent task success (versus retrieval metrics) is Not stated.

Problem: Deploying TTS in low-resource settings forces a tradeoff between expensive large voice-cloning models and compact fixed-voice systems that require a speaker-specific recorded corpus, and Thai adds further difficulty (ambiguous word boundaries, tone, code-switching, numeric verbalization).

Method: Distill a large zero-shot voice-cloning teacher (OmniVoice) into a compact 82M-parameter fixed-voice Kokoro-backbone student using entirely synthetic speech generated from a short (~15-second) voice reference, via a three-stage pipeline: Thai text preparation/verbalization (an LLM verbalizer rewrites ambiguous digits and embedded English into pronunciation-oriented Thai/Tinglish text), teacher sampling with quality filtering and rejection re-rendering (on pronunciation, pauses, rate, duration), and student training with a Thai–English phoneme frontend.

Key results:

• The resulting Wayu-Paxa-TTS-Edge model achieves 68.2% Challenge-Set Keyword Accuracy (85.5% of Gemini 3.1) and 91.4% pause precision — outperforming its own OmniVoice teacher's 89.9% pause precision and reaching 94.8% of Gemini 3.1.
• 3.7% CER on Thai and 1.1% CER on English.
• Best-of-K teacher sampling shows 15.1 percentage points of headroom (72.8% → 87.9% exact accuracy at K=118) not yet captured by the student.

Why it matters / caveats: Shows a practical "third route" for low-resource on-device TTS — pure synthetic distillation from a short reference, no speaker-specific recorded corpus needed — that beats its own teacher on some metrics. The paper notes synthetic voices generated by a teacher trained on in-the-wild speech could incidentally resemble real individuals, though the authors state they did not crawl or intentionally clone any real person's voice.

Problem: Building accurate OCR for lower-resource languages like Thai is hard because open OCR models/datasets/benchmarks concentrate on English and Chinese, real Thai documents have few reliable ground-truth labels, and it's unclear which properties of "realistic" synthetic training data actually drive transfer to real documents.

Method: A controlled document-reconstruction pipeline erases and re-renders source text regions in place (via HarfBuzz-based fit-constrained shaping) while independently varying source domain (Thai vs. translated-English), non-text page context, 2D layout, typeface diversity, and handwriting-glyph source, evaluated under both page-level and crop-level training using Qwen3-VL-2B-Instruct. Findings guide adaptation of the 0.9B-parameter PaddleOCR-VL-1.6 into Wayu-Paxa-OCR-Zero using 45,723 fully synthetic pages, with no real Thai OCR labels.

Key results:

• Non-text context has little consistent effect on transfer, while typeface diversity, 2D structure, and real handwriting glyphs help.
• In-domain reconstruction approaches real printed supervision under page-level training (1.82% vs. 1.31% median CER) but underperforms out-of-domain reconstruction under crop-level training (15.59% vs. 5.52%).
• Wayu-Paxa-OCR-Zero cuts median CER from 6.64% to 1.24% on printed pages and from 74.87% to 20.55% on handwriting relative to its base checkpoint, and outperforms Typhoon OCR v1 7B on all five evaluation sets.

Why it matters / caveats: Demonstrates that synthetic-only adaptation (no real OCR labels at all) can beat a larger real-data-trained open model (Typhoon OCR 7B) for a lower-resource language, and isolates which synthetic-data properties actually matter — useful methodology for extending OCR to other under-resourced scripts. The training-granularity-dependent domain-matching result (in-domain vs. out-of-domain reconstruction flipping which is best) suggests the recipe may need re-tuning for other languages.